Syntiant Corp. (500408) Earnings Call Transcript & Summary

Good morning, and welcome to this webinar on Trends and Solutions in AGI and TinyML, brought to you by Tata Elxsi and Syntiant thought leaders in this space, right? Before we start, let me do a quick request, let me make a quick request to our viewers. Can you please send in a quick question in the Q&A window so we know that we're receiving your questions. So when we have the segment -- Q&A segment at the end of the panel discussion, we will be able to see your questions, right? Thank you. So let me start by introducing our panelists. We have Mallik, Dr. Mallik Moturi. He's the Vice President of Business Development and Product Management of Syntiant. Mallik has been one of the first few early members of the leadership team of Syntiant. 30 years in the semicon business. He's a domain expert in AI. And Mallik, before I say anything else, let me say congratulations for the second CES Innovation Award that we received this year for the NDP120.

Thank you. Thank you, Bhaskar. Yes, appreciate it, and we are really honored to be in that -- this company of people who have been -- or companies who've been able to get that award. We did that last year as well. And the consumer electronic shows, as you know, has a -- but this year was -- we had to do it virtually, but we received a -- it was nice to have that. But yes, it's been a great run this whole 2020, though it's been unfortunate time for most people. We've been fortunate. Syntiant has actually had been able to get not just the award, but we also announced that we shipped more than 10 million of our first-generation product in the market already. So we have made our presence felt in the market this year. So thank you. Yes. So it's been interesting. I'm looking forward to this chat -- talk.

So are we. So are we. So here's fingers crossed for maybe a hat trick on the innovation awards, who knows next year.

Thank you.

Okay. You're most welcome. Let me introduce you to Dr. Milind Gandhe. Milind is the Vice President of the Systems Business unit and Smart Homes at Tata Elxsi. Many, many years in the company. And he and his team has been working on very exciting projects across the world. And Milind, I'm so glad you're on the panel, and I'm looking forward to a very interesting conversation.

I'm delighted to be here with the entire panel, Mallik, Naren and yourself Bhaskar. Just this morning, we announced our partnership with Syntiant. And very excited in terms of what this partnership can do for product -- for people who want to introduce products using a voice interface. Naren, do you want to introduce yourself?

Hey, hi. So thank you very much for this opportunity. And I think from a design point of view, especially when I sort of claim to represent the end consumers and end users of all the technology development and products that we are talking about, it's a great honor to be at this cutting-edge where we are actually talking about products that will revolutionize how people live their lives and what new things will happen in their homes and in the world around them. So thank you very much for giving me this opportunity. And yes, I'm with Tata Elxsi for the last 23 years, and I look at design.

Thank you.

Yes. So Mallik, while we wait for Bhaskar to come back, can I request you to give a very quick introduction to Syntiant?

Yes, definitely. Yes, as -- yes, no, I did not -- Hi, Bhaskar, welcome back. I think you had a small issue with webinex.

Sorry about that. Sorry about that.

So, no problem. So I know with all these interesting times we are living. So yes, we have a company which has been founded in 2017. Syntiant was founded in 2017 with our -- at that time, our lead investor being Intel Capital. And the founders, we have 4 founders, got together to start the company as an idea, a concept of at least one of our CEOs sitting with his the son, and they were -- he was just using the touchscreen and then later on started just talking to it. So he is like, yeah, this is really where we want to go. And thinking of -- we're all being in the semiconductor industry for a while, so we thought, oh, we should start a semiconductor company. And at those -- back in those days, everybody is, what are you doing? Or what -- how would you start a company in this? But interestingly enough, in the last few years, we've seen hundreds of AI chip companies come into being. We've been fortunate. We had not only great investors, we have a great team. We have about 70 employees right now. In the last 3 years, as I said earlier, we launched 2 products already in the market in production. The first generation one has already shipped more than 10 million chips. We -- our team is very experienced. Median experience in the company is more than 20 years. As I said, we've been very quick to market. We first built this -- first product was in 2018, production in 2019. And last year, 2020, our first -- our C round was completed with lead investors being Microsoft Ventures. So we have now not only Intel, we have Microsoft. We have Amazon Alexa Fund. We have Bosch and Atlantic Bridge joined us in the C round last year, and Applied Ventures. So we have really a good team of large number -- great investors with a lot of experience in the industry. And we are located in Irvine in Southern California. We have offices in India, Taiwan, China, U.K. and Denmark. So that's in a nutshell. Thank you.

[Technical Difficulty] I'm sorry, but I don't know that I missed out on introducing Naren. Did the introduction happen while I was away? I am not too sure. Naren, so let me introduce you to Naren -- Naren Ghate.

I have been introduced myself.

Yes, you did.

Yes, we made an introduction for Naren.

Perfect. It sounds good. so Naren, may I request you to give a quick introduction to Tata Elxsi?

Sure. So thanks so much Bhaskar. And Tata Elxsi is a member of the Tata Group, India's largest industrial conglomerate. The company focuses on product engineering services, but it's a -- the unique feature of the company is that it focuses not just on technology, but also on design. At more than 800-plus designers, we are probably South Asia's largest product design consultancy. And we believe that this sort of partnership between design and technology is what makes us a unique company. The company focuses on 4 verticals: automotive, broadcast, medical and smart home. What this knowledge of these verticals does -- so it's our belief, and we'll talk a little bit more about this today, is that -- the challenge that most product designers face is that they have to make sure that AI is not just a shiny new bubble, right, it's something that adds genuine value to their use -- to their customers. And that is possible only if the use case that AI enables is built with deep domain knowledge as well as tremendous empathy for users. And this is really the domain-specific customization that we hope to bring to the market along with Syntiant. So yes, looking forward, this is a very, very exciting times. The company has development centers in Bangalore, Pune, Chennai, Trivandrum and Mumbai in Southern India. And then also presence in North America, Europe, Japan as well as China and Singapore.

Yes, Bhaskar?

Thanks so much, Milind. So AI and the edge, ML, TinyML, these are all like buzzwords, right? These are very much realities now, right? So we have second-generation, third-generation products and technologies in this area. So very, very exciting times, right?

So my first one to you Mallik is how is AI at the edge revolutionizing technology products? And how it is enabling new and interesting use cases?

Thanks, Bhaskar. So we call it, as I said earlier, when we started the company, we're thinking about voice. But more, we are seeing the intelligence. We're seeing it, we call it the intelligence of things now. If you've heard of IoT, which is a broad category, but we are seeing a real renaissance in these technologies entering into these devices, right, like machines, automobiles and these things, which I said, we have been using that, which we use in our daily life, right, at home or office, factories. These are becoming more intelligent by adding neural networks or our intelligence of things what I call where basically that intelligence has been increasing in these devices over time. Because they simplify our life or enrich our users' lives by being intuitive, perceptive, so that you can take reasonable actions. So these devices can take reasonable actions on our behalf. For example, your smartphone is a very great example where you have set an alarm and you don't have to -- you can just talk to your device now. So this revolution is always the ability of AI to solve these problems more effectively than in the past. That's the revolution we are seeing. And one of the reasons why sudden there is this resurgence or this increase in AI in these edge devices is because the large amounts of data that's being generated, right? So as the data has been generated and collected, more efficient algorithms have been developed in the last few years. And because of the compute increase and the compute in the cloud, so we've been able to train more with the data that we have been available and the pipeline that is -- that researchers have developed more and more neural networks, more modeling techniques. Now that, combined with this also I heard I mentioned in the last few years, a number of companies would like ours, Syntiant has come up, but where you can bring, pull in that intelligence into the edge device in silicon. So a combination of silicon, data, and the algorithms to the pipeline is this revolution. That's the combination. Whoever has all of these magic components can put together a very good solution. And we have seen that. We are seeing that in lots of industries. And Syntiant itself, our first-generation product has so much acceptance. We've been hold into different sectors. We've been talking -- we're working with automobile companies and manufacturing, health, consumer, obviously, in consumer and retail. But let's look at, for example, just as an example of automobile as a revolution of technology, where products are I was just at my dealer -- car dealer recently, and he was asking -- I was just trying to ascertain the tire pressure monitoring. And the TPMS system there has a battery and that has to be replaced very frequently because it's got this -- sort of in target, but they still have to measure. But that's an example of place where starting -- we start seeing examples like TinyML showing up where we have very, very low power like Syntiant's device that could measure. And not only can it help you in the tire pressure monitor, but also on the road, while driving, adjusting the pressure as well as the information back to their user. So that's just an example that we are seeing that, and as I said, point of service in retail and hands-free or to -- for being hands free. We'll talk more in the next few minutes, but those are just generally, we're seeing health monitoring, agriculture and fisheries, manufacturing, a lot of different applications. Thank you.

Yes. And maybe I will give my point of view. Yes. I will give my point of view from the other side -- from side of the consumer. See in the consumer's mind, they'd like to quickly fall down to the simplest level of any device or any technology, any understanding that they have, so which means, like Mallik was saying, if a person's phone is intelligent, then very quickly, he's going to assume that the microwave and the washing machine and the oven, all of them can be intelligent, can be smart. And the other aspect of smartness that is changing very quickly and which will actually lead to the revolution that we're talking about, that a smart device is no longer just a connected device. I mean, you have a, let's say, televisions and other products that are connected to the Internet, but that will no longer be constituted as a smart device. A smart device would be something or somebody or a device, that can think for itself in its own location, in its own, let's say, geometry. And any device that has to sort of reach out to the cloud and talk to some other intelligent system and come back with an answer, will no longer be thought as a smart device. And that's where this entire TinyML and AGI really makes a huge step forward. Because in the perception of a consumer, if my device is smart on its own, then that's the true smart. And that's where we'll see a lot of demand coming in. What I expect happen is the market will start pulling for these kind of features in their products around them. And that's what would then push the industry to really deliver that level of experiences for the products. So yes, back to you. So Milind, maybe you could just add some of your points on this. Yes. Yes, Bhaskar, back to you sir.

Yes, I'm back. Sorry, I've been having terrible, terrible connection, so please bear with me. But those are fascinating -- it's fascinating inputs, Mallik, from you in intelligence of seeing the IoT, it's a new definition. And I love the example of the car OEM dealer. You obviously, have the name of the dealer. So I think we should go to market together, Tata Elxsi and Syntiant and make sure that OEM has TinyML in their TPMS system, yes. Naren, of course, what defines smart is very, very interesting. So my next question is actually to you, Mallik, is that we have seen a lot of interest and momentum in voice interface technology, especially given the current pandemic situation, and especially in consumer electronics, right? But do you also see interesting use cases beyond consumer electronics with voice? And my second part of the question is, what are the challenges and complexities in deploying AI-enabled the voice solutions?

Yes. Thanks, Bhaskar, and I'm sorry about your troubles though there, anyway. So yes, this pandemic situation is -- there's a lot of different use cases in different industries. Obviously, consumer electronics is a personal device. But in a public setting, we're seeing voice interface is actually quite effective, useful for hands-free operation. Let's take automobile, for example, or you're outside in a traveling, right. Hands-free control, for example, it's very important for automobile, right? It's very -- even not only for in this pandemic, but generally, hands-free control, where you're driving safety, driver safety, for vehicle entertainment or climate control, or vehicle access even to enter the automobile, like -- or you go to a shopping center, retail, you want -- at the point of service, if you can be hands-free, point-of-sale, I mean, shopping can be a quicker process, not only quicker, but also more effective with voice. So I'd see that as being a touchless, more touchless use cases everywhere. We are seeing that even in elevators, where you want to -- don't have to touch the buttons on the elevator. So I think the factory floor as well, we're seeing robots being controlled by voice. Doctors in operation or health care are looking at -- professionals there are looking at operating equipment with voice rather than -- because for one, they might -- hands might not be free. So I think access control with voice and other -- and in addition to all of this, including like if you want to enter a building with -- combined with voice, a face and voice biometrics are very useful. So we're seeing a lot of these kind of use cases in -- with voice-based control. We also see certain -- even mobile phones, mobile devices, obviously, now voice interface is really enabling a large number of people who do not have access to Internet previously to access it. I mean, if you think of it, voice is a much more natural interface. And literacy rates are not uniform across the world. So enabling voice-based interface to a mobile device is now -- we're seeing that with Google, with their Google Voice Assistant, we are seeing that across now other carriers also, their own interest, having -- adding their own voice interface to these devices. So we see this voice enabling a lot of people to enter into the Internet age. So that's what I think about this. Thank you.

No, thank you. Thank you. Very interesting use cases. Maybe I'll take a question to you, Milind, do you have any thoughts on this? Are these [indiscernible] technology in other industries, apart from consumer electronics?

Yes. So I think Mallik has painted a very broad picture in a very broad canvas. Let me just pick a couple of examples in going a little bit deeper. I think automotive, Mallik mentioned. And we definitely see infotainment as a major driver for voice interfaces. The dashboard, I think, I have always recognized that the dashboard is a major source of distraction for the driver, right? As the driver, you really want to keep your eyes on the road all the time. And seeking information on the dashboard, switching radio channels, picking music, et cetera, are all things that are taking your eyes off the road. The automotive industry did try to respond to that by moving controls to the steering wheel, and that does reduce distraction somewhat. But still, it does require you to take your eyes off the road for a little bit. So I think this is an absolutely natural sort of application for voice. And we see that coming in very, very strongly. We think that this will be more or less, the default sort of mode that you would interact with their infotainment system in no more than 2 or 3 years. On the other hand, in home appliances, we are already seeing voice, right? But the way we are seeing voice today is mostly in push to talk remotes, right? Whether it's an Alexa fire remote or whether it's an Apple TV remote, CD remote most of these remotes have a button that you push, and then you can say something and then your TV sort of response. But that's a somewhat unnatural way, right? And given that a lot of these appliances are pretty expensive and are always plugged in. So you don't have some of the BOM constraints as well as power constraints that you are facing. So at least my personal belief is that the time is right for far-field voice, where you're able to converse with your appliance up to 6 meters, that kind of sort of far-field voice applications will become commonplace in home appliances. Today, it's mostly TV, but I personally believe that air conditioners, room heaters, air purifies also sort of naturally sort of are designed for voice. Yes, that's really just my perspective on where voice is going. Bhaskar?

Thanks, Milind. I'm sorry, I don't know whether you can hear me, am I audible?

Yes.

No, I shut my video. So I'm hoping you're able to hear me. Yes. Are you? All right. All right. Thank you. Thank you. I'm sorry for the trouble.

Yes, yes, we can hear you.

So Milind, very, very interesting. Far-field voice applications, those are very, very interesting questions. Yes, very interesting areas. My next question is actually to Naren. And it goes, are we witnessing major changes in the way edge -- and can you share some insights on the design of UI?

So Milind painted this world where you can talk to devices around you. But imagine, if you don't know which devices can actually listen and which cannot, then as a consumer, you don't want to look dumb in front of your friends where you're talking to a toaster who is not smart, right? So what is going to happen is this would be a complete new semantics or new language of design that is going to come on the physical side of the product, so that you will know very naturally or very clearly that this is a device that is smart. This is a device that can understand voice instructions. And also it will direct you in terms of where to look and speak. Because as a person, as a human, you tend to look at the person that you're talking to. So you start looking at the device that you're talking to. And then it certainly means that the device is to acknowledge that you are looking at him, the device is to acknowledge that you are speaking to him, right? And so this automatically means that the way the products are designed, where it shows that it is accepting that kind of input, is seen in this design, whether it's the lights that come up or will be something else, right? So that is one aspect, which is the physical design of any of the products. The second, and I think a slightly more important part is the design of the voice interface itself. So one of the aspects of voice is -- as I guess, Mallik and Dr. Milind will agree, that if I start talking sophisticated English as a normal person, it takes a lot of compute power to really understand what I'm saying. And then even more compute power to really understand the nuances of what my meaning is, right? And we still make it really hard for the device to really correspond and respond to my instructions in the correct manner. So then if at this -- in the near short term, if I'm not able to really increase drastically the compute power of a device, can I do something so that the people actually start changing the way they talk to devices so that it becomes easier for the device to understand and it is natural or is expected for a human to sort of talk differently when talking to a device. I'll give a simple example of how to sort of get this change in behavior of a person for voice interface. So for example, if the voice that is responding back to you from the device or from the appliance, if that is very sophisticated, if that is very humane, that feels like a human person, the consumer is naturally going to respond in a human level of sophistication and expect the device behind voice to be as intelligence at the voice, right? So the voice sort of is a proxy for the intelligence of the device. On the other hand, if the voice sounds mechanical, sounds a bit robotic, then a human will naturally speak softly, maybe speak haltingly, maybe give very precise instructions. And that will make it easier for the device to understand what's happening, right? So just by changing how the voice sounds from the device, you can actually impact the behavior of the people using it and then make it easier for everybody to sort of have a good relationship with the product. Because we have to remember one thing, that in today's world, people don't give too many chances for a device to sort of -- they don't want to learn. They don't want to change too much about the way they are, and they expect the technology to catch up with them. But it's not that easy for the technology to catch up with human expectations, right? And if the human does not have enough of patience to really give microwave one more chance, then we have to sort of design the entire system such that the microwave has a chance to succeed and build a relationship with the human. The second point, and I'm just sort of continuing on this. The second example of what you could do from a voice interface design that will again help change and make this entire relationship succeed is to actually create this relationship. Now we talk about wake words. And many companies, whether it's Alexa or C, the reason why they use a single wake word is obviously, they want to create a single brand across the world where that would automatically be -- is something that people become used to. But going forward, if I'm able to give my microwave my own name, if I call my microwave Murphy and my neighbors' call sir, microwave Mickey, then what's going happen is, I want to give my poor microwave, poor Mickey some more leeway in learning. And I'm going to say that, oh, my poor Mickey is just 2 months old in my house is still learning to live with me. So I have to instruct it 2 or 3 time for it to understand what I say, right? So the moment I bring in these human emotions into this kind of relationship with my product, I'm giving a better chance for the product to succeed and better patients for the entire relationship to blossom. So those are simple tricks that we can use from a design point of view to really make some allowances for the technology that are needed. So those are some of my points on the physical and the voice interface design.

Thanks so much, Naren. I love the personification that you talked about. I think it's going to be really effective, right, if you can -- Mickey is in and -- thank you so much. So my next question is to you, Mallik. We've been seeing a huge surge in the adoption of TinyML, right, for AGI implementation. We'd love to hear your comments on that. What's triggering it?

Yes. I -- yes, thank you, again Bhaskar. So Naren's explanation of the voice UI and the design of UI is actually very interesting. I mean just if I may -- I'll get to that question in a second. It's -- one of the reasons why we find it very exciting to work with Tata Elxsi because of the design aspects and idea that we can work across. So we are a technology company, as you know, and this collaboration, we're looking forward to that kind of interaction. So let me talk about a little bit about the challenges. I mean, we are a technology company. We see challenges, right? I mean, one of the challenges, as Naren mentioned, is really in the problems -- the challenging problem of tackling voice interface just as speech itself is one of the more challenging of all the aspects of how humans interact. In the sense that it's a very diverse -- vocal tracks are very diverse. Everybody talks with a different accent, different parts of the world. I think it's even more challenging for AI than even image, which is a static image or you can detect what a object is. But here, it's a human talking, which have variation and voice, person to person, even the same person can sound different depending on their mood or time, location where they are in a noisy room, you might change your -- the level of your voice can be [ sound ]. So all of these things, even -- and even words many -- there are many homophones or homonyms in different words. So -- and you were saying, my poor microwave sits there, and it's supposed to recognize as well as a human being, that's a pretty challenging problem. And we are now telling all of this to be done in these TinyML devices, which are TinyML machine learning or Tiny devices with small amount of neurons. But if you can do it effectively at low power and performance like what we have, Syntiant, has been able to do, I think that's where TinyML is starting to kick off. And that's what we're seeing now why that resurgence or the growth in TinyML. I mean micro -- MCUs have been in the market now for many years. Today, the MCU market or microcontrolling market in 2020 was about 20 billion, 23 billion units per -- this year -- last year. We will grow to potentially about 30 billion by 2023. Now a large portion of this, we believe, and we are seeing that in the market and in fact, there are a lot of early adopters in this microcontroller in it, it's still in the early adoption stage with them. And we see -- we will see probably about 30% of that microcontrollers, which we see everywhere in our day-to-day life, 6 billion, 7 billion of those units in the next 5 years, will have machine learning or AI built into it, maybe even more. So we are seeing this growth because one of the challenges of TinyML as I was saying was they are very resource constrained, small power -- low power. So if you have to make it small and efficient and with lots of processing power because of just like an example of voice, one of the most challenging, again, as I was saying a few minutes ago, it's challenging. It has to be doing a lot of panel processing to do voice recognition. It's really a speech recognition engine. Though wake word, as Naren was mentioning, is one aspect of voice. Personalization. It could be like the microwave or the TV, for example, would respond differently for different members of the same family. So those are examples of where we are seeing this kind of personalization. You don't want to send all of this data. Privacy is a big factor. So we see TinyML ability to do -- reduce the latency, that means that you don't have to go to the cloud. Like, for example, you just -- you had this connection issue. That's an example of where you don't want to lose connectivity. And then suddenly, I cannot use my microwave because I can't have -- it doesn't have connection to the Internet. So that's robustness, and it has to be faster responding. You say stop, it has to stop. You cannot go to the network and cloud and then come back and say, "Oh, now you said stop, right? Then my popcorn will be burned by the time it stops. So that's what we're seeing. We're seeing a lot of this, as I said, high-performance engines like Syntiant's NDP being embedded inside these devices that -- and then the tools. There's a lot of consortiums also be a part of like machine learning commons, ML Commons. A number of big, large organized companies like Google and Facebook and Arm and Syntiant is part of it, Baidu as well. So around the world, a number of companies are coming together and providing tools and accelerators for machine learning. And I think that's what's driving this adoption of TinyML in the edge. Thanks. Bhaskar?

Yes. Thanks so much. And those numbers are huge, Mallik, when you talk about 30% or 30 billion potentially, those are huge numbers, absolutely. Milind, may I have your thoughts on this, please? On the surging adoption of TinyML, Milind?

It's perfect TinyML, really, right? To my mind, it is the coming together of 2 big trends. It's the coming together of IoT and it's coming together of AI. Now what's really driving this? And Mallik spoke a little bit from the user perspective. Let me also talk a little bit from the technology perspective, right? I think what's making this entire TinyMLs so common is really the coming together of 3 key trends: [ cheap ] links, a proliferation of sensors and low cost, low power, small compute elements at the edge. The kind of chips that Mallik has been talking about, right, the NDP100, the NDP200. And I'm saying there are sort of tiny compute engines only in comparison to what is available in the server. So if you go back, let's say, 3 or 4 years, a typical AI system would be architected primarily just call everything back to the cloud. Let the cloud do what it needs to do and come back. And then as Mallik said, the latency just goes out of the window. There's no real meaningful application. So there's impact on latency, there's impact on privacy. So the biggest sort of challenge, I think, for designers, for product designers is to really architect the system in such a way that you're able to trade-off what happens in the cloud and what happens at the edge. And a bunch of trade-offs, right? How much compute? How much memory? How much power? And how much bandwidth do you have available? So I think that's really the balancing act that we as designers need to keep in mind when we're designing some of these TinyML applications. The other, I think, key thing that also becomes more and more important with the proliferation of sensors is the need for sensor fusion at the edge. And again, it goes to the latency point that Mallik was referring to. And so when you're doing sensor fusion, you definitely need a basic amount of compute at the edge. And then it becomes very rapidly clear that you can't live with a simple microcontroller that you did when you were just doing a simple IoT system, right? So you need something that does a little bit more than that. And I think that's perhaps some of the places where NDP100 or NDP200 step into the breach. Yes. I think that's really where I think TinyML is going now.

Thank you so much for that, Milind. I know you mentioned sensor fusion. It's a favorite topic of Naren. I know we shall probably circle back to it if we have time. Very, very interesting. So my next question is for you, Mallik. Syntiant is playing such a key role in this TinyML and always on voice segment. So we'd love to hear more about Syntiant's road map moving forward, apart from the third award, of course, which we hope that you will get. But we'd love to hear more about your road map going forward.

Yes. Thank you, again, Bhaskar. So we -- I mean about -- talking about our products is always fun, especially I'm as responsible for products at Syntiant for me personally. But we start -- when we started developing the first generation, we call it the Syntiant core, the first-gen or Syntiant Core 1. So -- and then we have the -- with the NDP100, which Dr. Milind was mentioning about the NDP100, 101 and the -- those -- that family, it's around that technology. So these chips were -- they are currently in mass production. We have shipped more than, as I mentioned earlier, about 10 million to date. And we learned a lot with this first-gen neural decision processors. They are fully connected, feet forward neural networks with small. One of the advantages of being what we keeping it as the first generation and simple was that it was easy to program, get it into the market quickly and get market adoption. So we evolved that now for the gen 2, so we call it the Syntiant Core 2. And we have -- it's a remarkable new advance over the first gen. And it's almost 25x more in terms of the tensor throughput of the Syntiant Core 1. And it has a diversity of networks, which we can run on it, which can be programmed into it. And this NDP120, NDP200, as Naren was mentioning, 120 or 200 are our second next gen. And the 120 is the one that we just announced. It's built on Syntiant Core 2. We added a HiFi DSP on top of it as well as we have a microcontroller unit, a small MCU in there. But we envision the Syntiant Core 2 to be going beyond voice as well. And that's what we see as our -- what we see in our road map immediately. And then going beyond that into the next gen, we see Syntiant core -- next core after that to enable us to provide -- perform through ASR or natural language understanding our speech or a full automated speech recognition. And these will accomplish bigger and bigger tasks and still continue to be low power. Our core -- our 120, which we just launched is under a sub milliwatt we can do not just wake word, but we do a lot of different commands. We can do far-field, echo cancellation, and speaker verification. These are all things that today in the market, you would want an order of magnitude, maybe even 100 milliwatts of power, we can do it in sub milliwatt. So that's our road map. And I think we are looking forward to our customers adopting the Gen 2. And then beyond that, we will see how we can move into vision and image with our next-gen devices. And sensor fusion is definitely part of our sensor core in the sense that it can -- it's a concurrent network, so you can run voice as well as you can bring in other kinds of other -- if we have time, we can talk about the other sensors that we can pull in and connect and run neural networks on. Thanks.

Thank you, Mallik. Very, very exciting time. sub milliwatt power consumption that's phenomenal, absolutely phenomenal. You touched upon possible feature on your core 3, your next-gen processor as it were. And actually, it's going to address the next question I'm going to ask you, which is, is it really going to be possible to process natural language for richer conversation? That's number one. And number two is, regional language support as a -- do you see a big demand for that?

Yes. I mean, it is -- certainly is possible. I mean, it's being done today, right, richer conversation. But the question is, can it be done at edge -- at the edge at a really significantly low power, low-cost and affordably and a tiny enough to make it ubiquitous? Today, we are seeing that kind of richer conversation, talking about being hauled back to the cloud to do it. But we at Syntiant, we are implementing some modeling, acoustic models as well as more task-specific natural language models to run on these devices. Or at least at this stage, we are more in the exploratory phase. We are -- now we are experimenting and we are testing it. So on device AI, in the neural decision processors or NDPs will quickly determine the intent of the user. That's how we are thinking about it. So it will be specific to task or the use case that the customers want. So in a retail environment or travel, so that is a domain specific, and we can do a slot filling for -- depending on the user request. And then language models can then be added to do convert those models into -- for different languages. But those can tend to be too, if you want to cover all languages or those tend to be quite bulky in the language models for these devices today. And we think that those will be the part which will run in the cloud. But we do see dictation, speech to text, text to speech, real-time translation, that's things that we will be introducing in the next generation. And that's possible. So that's happening. That's what we are working on now. Thank you.

Yes. Thank you, Mallik. My next question, actually, I'm going to start that with a statement. We are very excited at Tata Elxsi at this partnership with Syntiant. I think it has tremendous potential. But from your perspective, Mallik, what benefits do you think this partnership brings to the competitive AGI market?

Yes. I mean, it's really interesting. I mean, when we first started talking last -- a couple of -- 2019 or last year, really, early last year. When we launch these products, we've got -- obviously, when we launched NDP, we had an overwhelming number of customers interested in our products. So when we -- a number of -- we pick, we would have to choose how many customers we could service or support. But when we started looking for a partner to work with, we were looking for someone who had ML and AI experience. But initially, we were looking at it as more for -- there's a lot of, as I was mentioning, for AI and ML, we need a lot of data, collection, modeling, and design. But data collection was where we thought we would work with, but then we quickly realized the design and like what Naren was saying, the entire design experience the VOI and then Milind's team on the AI side, we've discovered a lot more capabilities in Tata Elxsi that we could actually work much more closely with. So this is one of the reasons why we said, okay, we want to build this as a collaboration and then scale with the size of data, that we'll be able to work as a trusted partner that our customers will be then able to trust as well, and work with Tata and us. So this is the reason why we feel this is a good partnership, and we -- as Naren mentioned earlier, we just announced it in the -- it was -- there was -- I think there was some PR on this recently. But in general, we've been working very -- been partnering. And we did actually give a problem to Milind's team and early on and said, "Here is our evaluation kit, can you train pick your own wake word or commands and train it? And I think very, very quickly within 3 weeks or so, they were able to that -- your team was able to quickly bring together, collect the data and train it, which actually performed quite well. So we've said, this is definitely the right team to be working with. And also Tata has experience with other devices, smart devices, smart home, wearables, automobile -- automotive industry. So that's another area that we have a lot of customers talk to us about. So all these new verticals are very complementary to Syntiant. And we see this as an opportunity to collaborate with Tata Elxsi.

Thank you, Mallik. Great synergies. And as you can see, we do way more than data collection. So it's been serendipitous, and that means to discover each other. But I would love to hear from Milind what he feels about the partnership? And how excited we are about it?

Sure. Thanks, Bhaskar, and thanks, Mallik. So the way I look at it, right, I think Syntiant brings ultra-low power, always on voice solutions to the market. It's a platform. But to build a meaningful solution that a customer can use in the market, we need quite a few things on top of that platform. [indiscernible] natural and that is where Tata comes in. So the first thing you need to create a meaningful solution, I think, is domain knowledge, very, very important to create natural use cases -- understand the domain and create natural use cases. And I think that is -- the AI center of excellence at Tata Elxsi has been looking at needs of intelligence systems across domains, right? Whether it is automotive, whether it is consumer appliances, I think that the team has dealt with the challenges in multiple verticals. And dealt with situations where there are power constraints, dealt with situations where data is scarce. So a broad variety of challenges, I think, the team has dealt within. That way the experience in working with some of these domains, combined with the Syntiant's platform will allow us to create meaningful solutions. But I really want to go back and say that to pick on one thing that Mallik said, right? I think creating a meaningful AI solution is not just about AI, it's also about a lot of complementary stuff, right? So, a voice sensor is by its very nature in a noisy environment. You're not going to be in a sound-proof studio. And so a microphone is going to pick up all sorts of noise, right? And how do you deal with those noise? So there's a bunch of things you can do. You can do beam forming, you can do direction of arrival processing, you can do -- try and do things that improve the signal-to-noise ratio. And unless you do all of that, you're not going to be able to condition the signal well enough before it can be fed to a voice engine. So that is -- that sort of capability is very, very important. The other part, I think, is having the smart way of partitioning between the edge and the cloud and Mallik sort of alluded to this, language models today can be real beasts, right? If you look at something like GPT-3 or [indiscernible] you're looking at upwards of 70 billion parameters. And there's no way on earth that anybody is going to be able to fit that onto an edge computer. At the same time, if you put all of that in the cloud, then latency is going to go for a toss. So I think this partitioning between the edge and the cloud in the presence of these almost monster-sized language models, is something that is where a company like Tata comes in, in terms of helping the people plan the separation and live with that variety of mixed-use sort of AI applications.

Thanks, Milind. So bridging the gap is pretty much the exciting potential here. It's been a very interesting conversation. I want to move on to the Q&A section. We have a few really interesting questions that have come in, if I may. The first question is from Rajiv, and I'm going to read this out. It says AI has progressed to the extent that devices, understand the users' preferences at all times, especially for limited number of people, such as a family. Therefore, the aim of the U.S. designer should be to see how to eliminate UI and not complicated using multisensory interaction. What do you think? Naren, clearly, this one is for you. You are on mute, Naren.

Yes, sorry for that. Yes, it's a great question, and it's a raging debate in the design circles at least. I have a very firm opinion about it. And let me give a very simple example so that you understand where I'm coming from, especially when there's a discussion of whether UI should be needed in case of this hyper intelligent devices and environments around us. So let's take a car. And sometimes you want to be in this racing car where you can feel the power, you can feel the thrill when you're using the interface of your accelerator and the steering wheel and you're able to sort of really enjoy the experience. So here, the interface is something that you really want. Sometimes you want to be in this autonomous driving car, where it is taking you to your office and you want to just take a snooze, right? So the person is not a constant thing in the entire experience equation. So sometimes he wants to -- if he's cooking, he wants to feel that he's doing certain things by his hand. On the other hand, he want certain things that the devices understand automatically what he wants. So we should not take the person for granted. And that uncertainty about what a person wants and what kind of experiences he looks for is the real sort of future where there continues to be this nice balance that we're trying to achieve between what a person wants and how much intelligence should the device have. So it's partly how a UI might have to be designed for that.

Right. Right. Thanks, Naren, for that. Now another interesting question. I think we have time to take 2 or 3 more questions. So the next one that's come in, it says, is voice-biometric authentication the solution for user privacy and data security on NAS devices? Mallik, would you like to take that question?

Yes. As I was saying earlier, access or even what Naren was saying, user interface. So if you look at voice-biometric, there's a lot of variance to variation, various ways you can think of it. One is like speaker verification, right? Like, for example, is this the person who owns this device? So I have a microwave oven. I'm the only one who can tell it what to do or if I have a TV or I have Netflix channels, right? Your spouse and your children each, everybody has different channels, but just because of your voice, you could -- it can automatically know that you're the one who wants to watch the TV and then you can play that, right? That's a sort of speaker verification example. The other ones are more like -- it could also -- you want to false activate less. I was telling you about the challenges of voice. You don't want it to wake up just because somebody else said the wake word, right? So that is imposter rejection. That is another example of not -- it's a version of biometrics, it's a kind of biometer. If you want to go beyond that, like speaker identification, it becomes now, I want to enter -- unlock the car -- door of my car. I don't want -- only I need to. So that is a speaker, access control, it needs pretty strong because I don't want anybody random to open my home or my car or my bank account, which is yet another level of authentication and biometrics. So there are various ways you can break this up. But all of these come under the speaker understanding voice and vocal tracks and identifying who's speaking, right? So there's a lot of -- And that's all -- a lot of it is actually can be done neurally, and AI can effectively do that.

Very, very interesting thoughts on that. We will be able to take 2 quick questions. I'm going to read out the next one that's come in. It says benchmark is one of the challenges in TinyML. How are we addressing this? Milind, do you want to take that?

So, maybe, Mallik, you want to jump in first and I'll add on to that.

Sure. So -- yes, we are -- actually, benchmarking is a very important aspect. In fact, as I was telling earlier, we are a member -- our founding member actually of the ML Commons. And that's ML -- it is -- it's like a body or a lot of different companies have joined together to form ML Commons. One of the challenges is you need performance benchmarks. Otherwise, everybody -- without having benchmarks, you will be able -- nobody can come up with -- these are our -- there's no way to differentiate one product versus the other, and there's a lot of confusion in the market and the customers do not have a way. Today, even for voice, for example, Amazon did a good thing. They have actually have a benchmark, and you have to pass a certain -- a lot of other companies have picked on that and picked up -- picked that up and said, okay, these are the benchmark. So the idea is benchmarks are really critical because that can -- and if you have a common body that can come up with these sort of standards to say these are our metrics, and these are our performance benchmarks and then you meet these standards or these benchmarks. So that's actually what motivates the industry and the market and grows the market use cases and the customers. So you can actually then build devices also that meet those benchmarks. And at least as a performance reasons, right? So that's definitely a very critical aspect of machine learning.

Thanks for that, Mallik. We have time to take just one more question with a very quick answer, of course. So the question is, how is TinyML energy-efficient? How does it take care of meta data privacy by design, et cetera? I think there are 2 different questions, but, okay, let's take the first bit, which is how is TinyML is energy efficient?

Yes. Sure. I mean from a technology perspective, we -- that's how we are building devices, right? One of the -- how is it -- energy-efficient is, first of all, if you look at a microcontroller or a DSP, which is a Harvard architecture, what happens is you're fetching the instruction, you're fetching the data and then you run -- compute on that instruction and data, right? Instead of that, if you could pass the data locally. So one of our Syntiant's architecture is the data and compute are in the same -- they are in the -- on device. So in the sense that data compute is one aspect. The other one is parallel processing. Our machine learning requires -- it doesn't require like we have been used to in our CPUs and DSPs, 16, 32-bit and even 64 or 128-bit. On the other hand, a neural compute, 8-bit, or even 4-bit where like what we did in the first generation or even binary is more than sufficient to do very -- for neural compute. So those are the areas how we are reducing and improving our efficiency.

That certainly has been. And actually -- and some 1 milliwatt is what you mentioned, sometime like that itself kind of the goes through the way. Thank you so much. I think we're absolutely run out of time, but it's been a very, very interesting conversation. Thank you so much for joining. It's been fabulous. We must have another follow-up conversation soon. Thank you so much for joining. Thank you for your time, and thank you, participants and viewers who logged in. We hope to see you soon. Thank you, everybody. Have a great day, Mallik. Have a great evening, Naren and Milind.

Thank you. Good night, guys.

Thank you.

Thank you. Bye-bye. Bye.

Thank you.

Thank you.